SiamSYB: simple yet better methods to enhance Siamese tracking

Lixin Wei, Zeyu Xi, Ziyu Hu, Hao Sun

Abstract: Siamese-based single target trackers estimate the position of target in following frames of video. When facing complex scenes, obtaining accurate response map is the key to improve tracking performance. The robustness of most trackers is bad without template update. To solve these issues, a simple yet better tracking network (SiamSYB) is proposed. SiamSYB integrates the attention mechanism and template update module. With adding the attention mechanism, the network is more focus on the target. And the template update module makes network more robust when facing the challenges, including background clutter, similar objects and object deformation. Multi-stage offline training strategy is applied to get more accurate model. SiamSYB is the state-of-the-art tracker on 3 official test datasets, including VOT2016, VOT2019 and OTB100. SiamSYB achieves 0.391 EAO and 0.237 EAO on VOT2016 and VOT2019. SiamSYB achieves 0.853 precision score and 0.642 AUC score on OTB100. The tracking speed of SiamSYB is 90 FPS, which far surpasses the real-time speed of 25 FPS.